News

MuJoCo 4.0.1 corrects solver bug from major release

Google DeepMind's MuJoCo 4.0.1 fixes a sparse Cholesky factorization bug introduced in version 4.0.0 that could make the Newton solver converge slowly, produce incorrect accelerations or diverge.

D
Nov 19, 2026 · 2 min read

Google DeepMind released MuJoCo 4.0.1 to fix a sparse Cholesky factorization bug introduced in version 4.0.0. In models using a sparse constraint Jacobian, the project said, the defect could make its Newton solver converge slowly, calculate incorrect accelerations or diverge.

The fault gave the solver an incorrect Hessian factor, a mathematical representation used to choose each optimization step. It arrived with the major release five days earlier. The project’s release notes do not say how many users or models encountered it.

At issue was the logic for grouping adjacent matrix rows into a “supernode” so they could be processed together. The code treated two rows as equivalent when they shared the same columns and had the same number of descendant rows, even when the descendants themselves differed. According to the corrective commit, that mismatch could discard updates associated with one row and apply updates from another with the wrong coefficients.

The patch moves supernode detection into the symbolic factorization pass, checks the descendant sets there and supplies the resulting group sizes to the numeric factorization. It also adds a randomized regression test for sparse factor-and-solve operations across branches of an elimination tree.

The issue reporter who isolated the defect said a 92-degree-of-freedom hand-and-cube model showed acceleration differences of as much as 70% between sparse and dense paths on some steps. In that model, the sparse path hit the 100-iteration Newton limit, while the dense path needed no more than 11 iterations. The reporter said about half of slightly perturbed starting states ended with invalid or very large acceleration values.

Forcing the model to use the dense Jacobian removed the behavior, according to the reporter. Those model-specific measurements were not independently reproduced in the opened sources.

MuJoCo 4.0.0 arrived on November 6, five days before the patch. The maintainers called version 4 a major step under the project’s breaking-change versioning scheme, highlighting Studio and Filament tooling, deformables and speculative contacts in the major-release announcement.

Version 4.0.1 also brings the MuJoCo Studio Web Viewer to Windows and adds indexed reads for mjStringVec. It changes plugin loading so one unloadable dynamic library produces a warning instead of aborting simulate at startup, and it fixes multicontact problems in cylinder-edge collisions.

More news